Data Exploration - VET your beliefs

In Data Exploration we want to bootstrap our understanding of the dataset at hand. I like to think in terms of VET (Visualise-Explain-Test):

• Visualise: Wrangle your data and plot it in the most meaningful ways. Really, just get your hands dirty with it, plot everything.
• Explain: Given the patterns that appear in the data and your priors of the domain, can you build a hypothesis to explain the data?
• Test: This is the important part, cross-referencing your beliefs. Can you prove or disprove your hypothesis using additional data or following them to their logical conclusions?

This process takes the form of a loop; start with a dataset and a set of priors (beliefs) about that data, Visualise a new piece of information from the dataset, Explain the patterns by choosing a hypothesis, and Test it using an alternative dataset. If the Hypothesis is proved correct or incorrect, use it to update your beliefs. If it is not proved correct or incorrect, test it again.

In this way, given a reasonable dataset, you will bootstrap your priors into a more accurate or more confident set of beliefs.

Time Series Data - Lies, Damn Lies, and Cumulative Growth Statistics

It’s not immediately obvious, but cumulative growth statistics - for example annualised returns or compound annual growth rate - are the work of the devil.

There are a couple of things to really look out for when analysing

• Start Date: Why did the collector of this statistic pick the start date, is it different to other metrics?
• Survivorship Bias: What’s happened to examples of negative annualised returns? Are they being reported or have they been forgotten?